After deleting a node, its rowid and quantized vector remained in
other nodes' neighbor blobs via unidirectional reverse edges. This
is a data leak — the deleted vector's compressed representation was
still readable in shadow tables.
Fix: after deleting the node and repairing forward edges, scan all
remaining nodes and clear any neighbor slot that references the
deleted rowid. Uses a lightweight two-pass approach: first scan
reads only validity + neighbor_ids to find affected nodes, then
does full read/clear/write only for those nodes.
Tradeoff: O(N) scan per delete adds ~1ms/row at 10k vectors, ~10ms
at 100k. Recall and query latency are unaffected.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
diskann_repair_reverse_edges() ignored the return code from
diskann_node_write() when writing repaired neighbor lists after
a node deletion. A failed write would leave the graph inconsistent
with no error reported to the caller.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
NaN/Inf values in vectors break heap/sort invariants in KNN, causing
wrong or unpredictable results. Now rejected at parse time in
fvec_from_value() for both blob and JSON text input paths, with a
clear error message identifying the offending element index.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The rescore KNN loop read validity and rowids blobs from the chunks
iterator without checking their sizes matched chunk_size expectations.
A truncated or corrupt blob could cause OOB reads in bitmap_copy or
rowid array access. The flat KNN path already had these checks.
Adds corruption tests: truncated rowids blob and truncated validity
blob both produce errors instead of crashes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Infrastructure improvements:
- Fix benchmarks-ann Makefile: type=baseline -> type=vec0-flat (baseline
was never a valid INDEX_REGISTRY key)
- Add DiskANN + text primary key test: insert, KNN, delete, KNN
- Add rescore + text primary key test: insert, KNN, delete, KNN
- Add WAL concurrency test: reader sees snapshot isolation while
writer has an open transaction, KNN works on reader's snapshot
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Implements distance_hamming_avx2() which processes 32 bytes per
iteration using the standard VPSHUFB nibble-lookup popcount pattern.
Dispatched when SQLITE_VEC_ENABLE_AVX is defined and input >= 32
bytes. Falls back to u64 scalar or u8 byte-at-a-time for smaller
inputs.
Also adds -mavx2 flag to Makefile for x86-64 targets alongside
existing -mavx.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New benchmarks-ann/bench-delete/ directory measures KNN recall
degradation after random row deletion across index types (flat,
rescore, IVF, DiskANN). For each config and delete percentage:
builds index, measures baseline recall, copies DB, deletes random
rows, measures post-delete recall, VACUUMs and records size savings.
Includes Makefile targets, self-contained smoke test with synthetic
data, and results DB for analysis.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
DiskANN's delete repair only fixes forward edges (nodes the deleted
node pointed to). Stale reverse edges can cause deleted rowids to
appear in search results. Fix: track a 'confirmed' flag on each
search candidate, set when the full-precision vector is successfully
read during re-ranking. Only confirmed candidates are included in
output. Zero additional SQL queries — piggybacks on the existing
re-rank vector read.
Also adds delete hardening tests:
- Rescore: interleaved delete+KNN, rowid_in after deletes, full
delete+reinsert cycle
- DiskANN: delete+reinsert cycles with KNN verification, interleaved
delete+KNN
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Casting unaligned blob pointers to u64* is undefined behavior on
strict-alignment architectures. Use memcpy to safely load u64 values
from potentially unaligned memory (compilers optimize this to native
loads on architectures that support unaligned access).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
vmulq_s16(diff, diff) produced int16 results, but diff can be up to
255 for int8 vectors (-128 vs 127), and 255^2 = 65025 overflows
int16 (max 32767). This caused NaN/wrong results for int8 vectors
with large differences.
Fix: use vmull_s16 (widening multiply) to produce int32 results
directly, avoiding the intermediate int16 overflow.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The binary quantizer uses D/8 for buffer sizes and memset, which
truncates for non-multiple-of-8 dimensions, causing OOB writes.
Rather than using ceiling division, enforce the constraint at
table creation time with a clear parse error.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The `dist` variable in rescore KNN quantized distance computation was
uninitialized. If the switch on quantizer_type or distance_metric
didn't match any case, the uninitialized value would propagate into
the top-k heap, potentially returning garbage results.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
sqlite3_column_blob() returns NULL for zero-length blobs or on OOM.
Several call sites in rescore KNN and DiskANN node/vector read passed
the result directly to memcpy without checking, risking NULL deref on
corrupt or empty databases. IVF already had proper NULL checks.
Adds corruption regression tests that truncate shadow table blobs and
verify the query errors cleanly instead of crashing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
vec0Update_UpdateVectorColumn writes to flat chunk blobs but does not
update DiskANN graph or IVF index structures, silently corrupting KNN
results. Now returns a clear error for these index types. Rescore
UPDATE is unaffected — it already has a full implementation that
updates both quantized chunks and float vectors.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Six sites used #if SQLITE_VEC_ENABLE_RESCORE to guard _vector_chunks
skip logic that applies to ALL non-flat index types. When RESCORE was
compiled out, DiskANN and IVF columns would incorrectly access flat
chunk tables. Two sites also missed DiskANN in the skip enumeration,
which would break mixed flat+DiskANN tables.
Fix: replace all six compile-time guards with unconditional runtime
`!= VEC0_INDEX_TYPE_FLAT` checks. Also move rescore_on_delete inside
the !vec0_all_columns_diskann guard to prevent use of uninitialized
chunk_id/chunk_offset, and initialize those variables to 0.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The UINT32_TYPE/UINT16_TYPE/INT16_TYPE/UINT8_TYPE/INT8_TYPE/LONGDOUBLE_TYPE
macros (copied from sqlite3.c) were never used anywhere in sqlite-vec.
The u_int8_t/u_int16_t/u_int64_t typedefs redefined standard types using
BSD-only types despite <stdint.h> already being included, breaking builds
on musl/Alpine, strict C99, and requiring reactive platform guards.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extend benchmarks-ann/ with results database (SQLite with per-query detail
and continuous writes), dataset subfolder organization, --subset-size and
--warmup options. Supports systematic comparison across flat, rescore, IVF,
and DiskANN index types.
The command insert handler (used for runtime config like
search_list_size_search) was gated behind SQLITE_VEC_EXPERIMENTAL_IVF_ENABLE,
which defaults to 0. DiskANN commands were unreachable unless IVF was
also compiled in. Widen the guard to also activate when
SQLITE_VEC_ENABLE_DISKANN is set.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add DiskANN graph-based index: builds a Vamana graph with configurable R
(max degree) and L (search list size, separate for insert/query), supports
int8 quantization with rescore, lazy reverse-edge replacement, pre-quantized
query optimization, and insert buffer reuse. Includes shadow table management,
delete support, KNN integration, compile flag (SQLITE_VEC_ENABLE_DISKANN),
release-demo workflow, fuzz targets, and tests. Fixes rescore int8
quantization bug.
Rename SQLITE_VEC_ENABLE_IVF to SQLITE_VEC_EXPERIMENTAL_IVF_ENABLE and
flip the default from 1 to 0. IVF tests are automatically skipped when
the build flag is not set.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add inverted file (IVF) index type: partitions vectors into clusters via
k-means, quantizes to int8, and scans only the nearest nprobe partitions at
query time. Includes shadow table management, insert/delete, KNN integration,
compile flag (SQLITE_VEC_ENABLE_IVF), fuzz targets, and tests. Removes
superseded ivf-benchmarks/ directory.
The Linux AVX auto-detection checked the host's /proc/cpuinfo, which
passes on x86 CI runners even when cross-compiling for Android ARM
targets. Skip AVX detection when CC contains 'android'.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add approximate nearest neighbor infrastructure to vec0: shared distance
dispatch (vec0_distance_full), flat index type with parser, NEON-optimized
cosine/Hamming for float32/int8, amalgamation script, and benchmark suite
(benchmarks-ann/) with ground-truth generation and profiling tools. Remove
unused vec_npy_each/vec_static_blobs code, fix missing stdint.h include.
vec0Update_Delete_ClearMetadata's long-text branch runs a DELETE via
sqlite3_step, which returns SQLITE_DONE (101) on success. The code
checked for failure but never normalized the success case to SQLITE_OK.
The function's epilogue returned SQLITE_DONE as-is, which the caller
(vec0Update_Delete) treated as an error, aborting the DELETE scan and
silently leaving rows behind.
- Normalize rc to SQLITE_OK after successful sqlite3_step in ClearMetadata
- Move sqlite3_finalize before the rc check (cleanup on all paths)
- Add test_delete_by_metadata_with_long_text regression test
- Update test_deletes snapshot (row 3 now correctly deleted)
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add rescore index type: stores full-precision float vectors in a rowid-keyed
shadow table, quantizes to int8 for fast initial scan, then rescores top
candidates with original vectors. Includes config parser, shadow table
management, insert/delete support, KNN integration, compile flag
(SQLITE_VEC_ENABLE_RESCORE), fuzz targets, and tests.
Add approximate nearest neighbor infrastructure to vec0: shared distance
dispatch (vec0_distance_full), flat index type with parser, NEON-optimized
cosine/Hamming for float32/int8, amalgamation script, and benchmark suite
(benchmarks-ann/) with ground-truth generation and profiling tools. Remove
unused vec_npy_each/vec_static_blobs code, fix missing stdint.h include.
When a row is deleted from a vec0 virtual table, the rowid slot in
_chunks.rowids and vector data in _vector_chunksNN.vectors are now
zeroed out (previously left as stale data, tracked in #54). When all
rows in a chunk are deleted (validity bitmap all zeros), the chunk and
its associated vector/metadata shadow table rows are reclaimed.
- Add vec0Update_Delete_ClearRowid to zero the rowid blob slot
- Add vec0Update_Delete_ClearVectors to zero all vector blob slots
- Add vec0Update_Delete_DeleteChunkIfEmpty to detect and delete
fully-empty chunks from _chunks, _vector_chunksNN, _metadatachunksNN
- Fix missing rc check in ClearMetadata loop (bug: errors were silently
ignored)
- Fix vec0_new_chunk to explicitly set _rowid_ on shadow table INSERTs
(SHADOW_TABLE_ROWID_QUIRK: "rowid PRIMARY KEY" without INTEGER type
is not a true rowid alias, causing blob_open failures after chunk
delete+recreate cycles)
- Add 13 new tests covering rowid/vector zeroing, chunk reclamation,
metadata/auxiliary/partition/text-PK/int8/bit variants, and
page_count shrinkage verification
- Add vec0-delete-completeness fuzz target
- Update snapshots for new delete zeroing behavior
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Homebrew LLVM 18 runtime dylibs use typed allocation ABI symbols
(__ZnwmSt19__type_descriptor_t) not available in macOS 14's system
libc++, causing dyld to abort. Xcode clang doesn't ship libFuzzer.
Mark fuzz-macos as continue-on-error (same as fuzz-windows) so it
doesn't block CI. Linux fuzzing remains the primary bug detector.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
sqlite-vec.c:
- vec0_free: add loops to free partition, auxiliary, and metadata
column names (previously leaked on error paths)
- vec0_init: update pNew->numXxxColumns incrementally in the parse
loop so vec0_free sees correct counts on early goto-error paths
(previously the counts were only written after the loop, so vec0_free
would loop 0 times and leak names allocated inside the loop)
fuzz.yaml:
- macOS: pass -isysroot $(xcrun --sdk macosx --show-sdk-path) so
Xcode clang can find system headers (stdio.h, assert.h, etc.)
- Fix artifact upload paths: libFuzzer writes crash-*/leak-* to
the cwd (repo root), not tests/fuzz/
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Homebrew LLVM 18's ASAN/libFuzzer runtimes contain weak-def symbols
for typed allocation operators (__ZnwmSt19__type_descriptor_t) that
don't exist in macOS 14's system libc++. Since the symbol is embedded
in the pre-built runtime dylibs (not our code), link-time flags cannot
fix it.
Switch to Apple's Xcode clang which ships its own libFuzzer and ASAN
runtime built against the system libc++ — no ABI mismatch possible.
No Homebrew LLVM install needed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The fuzz targets were crashing on macOS 14 with:
dyld: weak-def symbol not found '__ZnwmSt19__type_descriptor_t'
libFuzzer compiled with LLVM 18 uses typed allocation ABI symbols
not present in macOS 14's system libc++. Since DYLD_LIBRARY_PATH
cannot override SIP-protected /usr/lib/libc++.1.dylib at runtime,
we fix this at link time:
- -nostdlib++: suppress implicit system libc++ linking
- -L$LLVM/lib/c++ -lc++: explicitly link LLVM's libc++ (which has the symbol)
- -Wl,-rpath,$LLVM/lib/c++: embed rpath so dyld finds LLVM's libc++ at runtime
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
